通过最小化最大熵，实现数据高效且可控的生成式AI
A new pathway to generative artificial intelligence by minimizing the maximum entropy
- 不拟合训练数据，而是寻找信息量高且噪声低的数据表示
- 在有限数据下生成图像效果优于变分自编码器，支持后验定制
- 适合需要灵活控制生成过程的研究者与应用开发者
生成式人工智能已深刻改变社会。当前模型通过最小化生成数据与训练集之间的距离进行训练，导致发展趋于瓶颈：模型高度依赖数据，且生成过程难以控制。为突破这一限制，本文提出新范式：不拟合训练集，而是同时寻找最具信息量且最无噪声的数据表示，通过对抗训练最小化熵以减少噪声，最大化熵以保持无偏。由此得到一种基于物理规律的通用模型，具备数据高效性和灵活性，可对生成过程进行控制和干预。基准测试表明，该方法优于变分自编码器；即使在少量训练数据下也能生成高质量图像，并实现无需微调或重训练的生成过程后验定制。
原文摘要 · Abstract (English)
Generative artificial intelligence revolutionized society. Current models are trained by minimizing the distance between the produced data and the training set. Consequently, development is plateauing as they are intrinsically data-hungry and challenging to direct during the generative process. To overcome these limitations, we introduce a paradigm shift through a framework where we do not fit the training set but find the most informative yet least noisy representation of the data simultaneously minimizing the entropy to reduce noise and maximizing it to remain unbiased via adversary training. The result is a general physics-driven model, which is data-efficient and flexible, permitting to control and influence the generative process. Benchmarking shows that our approach outperforms variational autoencoders. We demonstrate the methods effectiveness in generating images, even with limited training data, and its unprecedented capability to customize the generation process a posteriori without any fine-tuning or retraining
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。